RESEARCH ARTICLE10.1002/2015JC011283

Short-term sea ice forecasting: An assessment of iceconcentration and ice drift forecasts using the U.S. Navy’sArctic Cap Nowcast/Forecast SystemDavid A. Hebert1, Richard A. Allard1, E. Joseph Metzger1, Pamela G. Posey1, Ruth H. Preller1,Alan J. Wallcraft1, Michael W. Phelps2, and Ole Martin Smedstad3

1Naval Research Laboratory, Stennis Space Center, Mississippi, USA, 2Jacobs Technology Inc., Stennis Space Center,Mississippi, USA, 3Vencore Services and Solutions, Inc., Stennis Space Center, Mississippi, USA

Abstract In this study the forecast skill of the U.S. Navy operational Arctic sea ice forecast system, theArctic Cap Nowcast/Forecast System (ACNFS), is presented for the period February 2014 to June 2015.ACNFS is designed to provide short term, 1–7 day forecasts of Arctic sea ice and ocean conditions. Manyquantities are forecast by ACNFS; the most commonly used include ice concentration, ice thickness, icevelocity, sea surface temperature, sea surface salinity, and sea surface velocities. Ice concentration forecastskill is compared to a persistent ice state and historical sea ice climatology. Skill scores are focused on areaswhere ice concentration changes by 65% or more, and are therefore limited to primarily the marginal icezone. We demonstrate that ACNFS forecasts are skilful compared to assuming a persistent ice state, espe-cially beyond 24 h. ACNFS is also shown to be particularly skilful compared to a climatologic state for fore-casts up to 102 h. Modeled ice drift velocity is compared to observed buoy data from the InternationalArctic Buoy Programme. A seasonal bias is shown where ACNFS is slower than IABP velocity in the summermonths and faster in the winter months. In February 2015, ACNFS began to assimilate a blended ice concen-tration derived from Advanced Microwave Scanning Radiometer 2 (AMSR2) and the Interactive MultisensorSnow and Ice Mapping System (IMS). Preliminary results show that assimilating AMSR2 blended with IMSimproves the short-term forecast skill and ice edge location compared to the independently derivedNational Ice Center Ice Edge product.

1. Introduction

Recent satellite observations have shown that the Arctic September sea ice extent minimum has beendecreasing at a more rapid rate in the past two decades (1997–2014) than the first two decades when satel-lite observations were first made in 1979 [Serreze and Stroeve, 2015, Figure 2]. This increased rate of declinein sea ice cover means a greater area of navigable waters in the Arctic will be available for longer periods oftime. The ability to forecast ice conditions is of crucial importance for maritime operational planning [U.S.Navy, 2014] and for scientific applications such as monitoring oceanographic [Johannessen et al., 2004;Serreze et al., 2007] and biologic [Arrigo et al., 2008; Aschan and Ingvaldsen, 2009; Durner et al., 2011; Jayet al., 2011] conditions. The Naval Research Laboratory (NRL) has a long history of Arctic forecast systems,going back to 1987 with the Polar Ice Prediction System (PIPS) [Preller et al., 2002], an Arctic basin forecastsystem consisting of the Hibler ice model [Hibler, 1979, 1980] and a monthly mean ocean climatological forc-ing. This system was replaced by PIPS 2.0, a 25 km resolution system coupling the Hibler ice model to theCox ocean model [Cox, 1984] and encompassed all the sea ice covered regions of the northern hemisphere.With many recent improvements in sea ice and ocean modeling, as well as advances in high performancecomputing, the next generation Arctic forecast system, the Arctic Cap Nowcast/Forecast System (ACNFS),has been developed. ACNFS uses the Los Alamos Community Ice CodE (CICE) version 4.0 [Hunke andLipscomb, 2008] as the sea ice model which is two-way coupled to the HYbrid Coordinate Ocean Model(HYCOM) [Bleck, 2002; Metzger et al., 2014, 2015].

There are many ice-ocean modeling systems currently in use. The Community Earth System Model (CESM)[Kay et al., 2015] is a global, fully coupled atmospheric-ice-ocean model and is freely available at https://www2.cesm.ucar.edu/models/current. The Estimating Circulation and Climate of the Ocean (ECCO) project

Special Section:Forum for Arctic Modelingand Observing Synthesis(FAMOS): Results and Synthe-sis of CoordinatedExperiments

Key Points:� Five day ice concentration forecast

skill, ice edge prediction and icemotion compared to observations� Improve forecasts by assimilating

National Ice Center IMS into U.S. Navysea ice forecast system� Skill of U.S. Navy ice forecasting

compared to persistence andclimatology

Correspondence to:D. Hebert,david.hebert@nrlssc.navy.mil

Citation:Hebert, D. A., R. A. Allard, E. J. Metzger,P. G. Posey, R. H. Preller, A. J. Wallcraft,M. W. Phelps, and O. M. Smedstad(2015), Short-term sea ice forecasting:An assessment of ice concentrationand ice drift forecasts using the U.S.Navy’s Arctic Cap Nowcast/ForecastSystem, J. Geophys. Res. Oceans, 120,doi:10.1002/2015JC011283.

Received 1 SEP 2015

Accepted 19 NOV 2015

Accepted article online 23 NOV 2015

VC 2015. American Geophysical Union.

All Rights Reserved.

Published 2015. This article is a U.S.

Government work and is in the public

domain in the USA.

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 1

Journal of Geophysical Research: Oceans

PUBLICATIONS

is a global ice-ocean state estimation using an adjoint method [Stammer et al., 2002; Wunsch and Heimbach,2013]. The Canadian Global Ice Ocean Prediction System (GIOPS) [Smith et al., 2015] provides a daily global iceand ocean analysis and 10 day forecasts using CICE coupled to the eddy-permitting, 1/48 Nucleus for Euro-pean Modeling of the Ocean (NEMO). Several regional ice-ocean models are also published. Full Arctic domainmodels include the Pan Arctic Ice Ocean Modeling and Assimilation System (PIOMAS) [Schweiger et al., 2011;Zhang and Rothrock, 2003], a recent evolution of PIOMAS called Marginal Ice Zone Modeling and AssimilationSystem (MIZMAS), [Schweiger and Zhang, 2015], the Naval Postgraduate School Arctic Modeling Effort (NAME)[Maslowski et al., 2004; McGeehan and Maslowski, 2012], and the TOPAZ System [Bertino and Lisæter, 2008;Sakov et al., 2012]. Smaller sub-arctic regions include the Canadian Regional Ice Prediction System (RIPS)[Buehner et al., 2013; Lemieux et al., 2015], the Nares Strait [Rasmussen et al., 2010], and Labrador Sea andBaffin Bay [Fenty and Heimbach, 2013]. Finite element [Scholz et al., 2013; van Scheltinga et al., 2010] and finitevolume [Gao et al., 2011] ice models also have been used to study ice-ocean dynamics. With the exceptionof the Canadian RIPS, MIZMAS, and TOPAZ, these models are used for either hindcasts (predictions of pastevents) or multi-decadal climate studies. ACNFS, in contrast, is designed to produce short-term, 1-7 day fore-casts of current ice conditions with a focus on the marginal ice zone where ships are most likely to travel.

The aim of this paper is to document ACNFS and validate sea ice forecasts up to 5 days. A detailed descrip-tion of the model is presented in section 2, followed by the forecast assessments and results in section 3. Asummary is provided in section 4.

2. The Arctic Cap Nowcast/Forecast System (ACNFS)

ACNFS is a data assimilative, two-way coupled sea ice – ocean model driven by external atmospheric forc-ing. ACNFS [Posey et al., 2010] was declared operational by the U.S. Navy in September 2013, and is run dailyproviding Arctic sea ice forecasts out to 7 days. A description of each model, coupling between models,forcing, and data assimilation are now presented.

2.1. Ice Model: CICE V4.0The Los Alamos Community Ice Code (CICE) version 4.0 [Hunke and Lipscomb, 2008] is the ice componentcurrently used in ACNFS. Within CICE, sea ice and snow are divided into several discrete ice thickness cate-gories. In ACNFS there are five thickness categories; within each thickness category there are four ice layersand one snow layer (these settings are the CICE defaults). For each ice and snow layer, changes in thicknessby thermodynamic processes including radiative, turbulent, and conductive heat fluxes are computed viaan energy conserving thermodynamic model [Bitz and Lipscomb, 1999]. The model ice thickness distributionvaries in time due to thermodynamic and mechanical processes (e.g., ridging), and is updated using the icethickness remapping scheme of Lipscomb [2001]. Ice momentum is solved using the elastic-viscous-plastic(EVP) model [Hunke and Dukowicz, 1997, 2002]. EVP is a modification to traditional viscous-plastic (VP) mod-els [e.g., Hibler, 1979] where elastic waves are allowed to exist in the ice. This modification allows themomentum equation to be computed with a much more efficient, fully explicit numerical scheme, asopposed to an implicit scheme used in VP models. The elastic waves, while nonphysical, are damped out bysubcycling every time step. In ACNFS the time step is 10 min, with 120 subcycles per time step.

2.2. Ocean Model: HYCOMThe HYbrid Coordinate Ocean Model (HYCOM) [Bleck, 2002; Chassignet et al., 2003] is the ocean model usedin ACNFS. HYCOM combines three different vertical coordinates: (1) pressure (fixed depth), best in mixedlayer and unstratified ocean, (2) sigma or terrain following, often the best choice in shallow water, and (3)isopycnal (density), with best application in the deep stratified ocean. There are 32 HYCOM vertical layers inACNFS. While the number of total layers stays the same, it is possible for the number of each type of layerto change each time step. The model makes a dynamically smooth transition between coordinate types byusing a layered continuity equation. The K-Profile Parameterization (KPP) [Large et al., 1994] is used toparameterize vertical mixing.

2.3. Ice-Ocean Model CouplingIn ACNFS, CICE and HYCOM are two-way coupled via the Earth System Modeling Framework (ESMF)[Campbell et al., 2010; Hill et al., 2004]. CICE and HYCOM are run individually and exchange output everymodel hour. Table 1 contains a list of quantities coupled between each model. Ice thickness, fluxes, and

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 2

stresses are sent to HYCOM from CICE while seasurface temperature, sea surface salinity, and seasurface currents are sent from HYCOM to CICE.

2.4. Atmospheric ForcingThe atmospheric forcing for ACNFS is currentlyprovided by the NAVy Global Environment Model(NAVGEM) [Hogan et al., 2014]. (Prior to August2013, forcing was provided by the Navy Opera-

tional Global Atmospheric Prediction System (NOGAPS) [Hogan et al., 1991]. NOGAPS was decommissionedin August 2013, with NAVGEM taking its place. Further details about the conversion from NOGAPS to NAV-GEM are provided in Metzger et al. [2013].) NAVGEM is a spectral model with triangular truncation at wavenumber 425 (approximately 0.28 degree horizontal resolution). Output from NAVGEM is currently providedfor use by ACNFS and other Navy applications on a uniformly spaced, 0.5 degree grid. (In the future, NAV-GEM output is planned to be provided on its native 0.28 degree grid). NAVGEM provides output fields at 3 hintervals including 2 m air temperature, surface humidity, net surface shortwave and longwave radiation,precipitation, 10 m zonal and meridional wind velocities, mean sea level pressure, and 2 m dew point. Thesevalues are interpolated to the ACNFS computational grid for use in each daily run.

2.5. Data AssimilationObservational data assimilation is achieved via the Navy Coupled Ocean Data Assimilation (NCODA)[Cummings, 2011; Cummings and Smedstad, 2013]. Data assimilation is performed to reduce errors in modelforecasts due to several factors including nonlinear processes that are not a deterministic response toatmospheric forcing, poorly parameterized physical processes, limitations in numerical algorithms, and limi-tations due to model resolution. NCODA uses the 3-D variational analysis (3DVAR) method to assimilateobservations into the model. Observational data that are assimilated into ACNFS come from a variety ofsources. Most of the observed ocean surface data come from satellite-borne instruments. Sea surface heightis obtained from Ocean Surface Topography Mission (OSTM)/Jason-2, AltiKa, and CryoSat-2. Sea surfacetemperature is obtained from the Advanced High Resolution Radiometer (AVHRR) (AVHRR methodology isthe same as described in Cayula et al. [2013] for Visible Infrared Imaging Radiometer Suite (VIIRS)), Geosta-tionary Operational Environmental Satellite (GOES), Meteosat Second Generation (MSG) satellite, andAdvanced Microwave Scanning Radiometer 2 (AMSR2). Ice Concentration is obtained from the DefenseMeteorological Satellite Program (DMSP) Special Sensor Microwave Imager/Sounder (SSMI/S) and AMSR2.Ice concentration from SSMI/S is derived using the NASA Team 2 sea ice algorithm [Markus and Cavalieri,2000, 2009], while ice concentration from AMSR2 is derived from the Bootstrap Algorithm [Comiso, 1986;Comiso and Nishio, 2008]. While these observations are useful for correcting surface conditions, observationsbelow the surface are also desired to help correct ocean quantities at depth. Observations below the surfacethat are assimilated into ACNFS include temperature and salinity profiles from expendable bathy-thermographs (XBT) and conductivity-temperature-depth (CTD) profiles. Ice-Tethered Profiler data [Krishfieldet al., 2008; Toole et al., 2011] provide temperature and salinity profiles below the sea ice to depths of 760meters, and are available in near real-time and assimilated into HYCOM. Sea surface height observations areconverted to a synthetic vertical temperature and salinity profile via the Modular Ocean Data AssimilationSystem (MODAS) system [Fox et al., 2002]. Temperature observations are also obtained from profiling floatssuch as Argo (www.argo.net).

NCODA performs data assimilation in a two-step process. First, all observational data are quality controlled(QC’d) to ensure erroneous data are not assimilated. The QC process is detailed in Cummings [2005, 2011].Briefly, a series of initial ‘‘sanity checks’’ are performed including a land-sea boundary test and valid physicaldata range tests, followed by instrumentation error checks, and then cross validation checks for consistencyof observations. After these checks, the data are compared to ‘‘background’’ fields, which include climatol-ogy and short-term (past few weeks) observation analyses. A probability of error is then calculated basedon the difference between the observation and background data. The sea ice concentration climatology isobtained from the European Centre for Medium Weather Forecast (ECMWF) climatology formed from years1979 to 1996 [Fernandez et al., 1998], while ocean climatologies are obtained from Generalized Digital Envi-ronmental Model (GDEM) [Carnes et al., 2010]. If the observation probability of error reaches a specified

Table 1. Quantities Exchanged Between CICE and HYCOM EveryHour

From CICE to HYCOM From HYCOM to CICE

Ice concentration Sea surface temperatureIce x/y stresses Sea surface salinityHeat flux through ice Sea surface x/y velocitiesIce freeze/melt heat fluxNet water flux

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 3

threshold (usually 5%), the data are not assimilated. Second, the QC’d data are then used in the 3DVARmethod to create an analysis field based on the observed and background data and the error covariance.The NCODA ice and ocean analysis fields are then used for data assimilation into the model.

2.5.1. Assimilation With IMSThe Interactive Multisensor Snow and Ice Mapping System (IMS) [Helfrich et al., 2007] is an operational snowand ice mask product produced daily and valid at 00Z (Zulu, or Greenwich Mean Time). IMS is produced byan analyst at the National Ice Center (NIC) and incorporates a multitude of satellite imagery data includingvisible/infrared (VIS/IR), synthetic aperture radar (SAR), scatterometer, and passive microwave. The result is asea ice extent mask that is available at either 1 km or 4 km spatial resolution. The IMS ice mask is availableas a plain text file and GeoTiff [NIC, 2008]. (Note that IMS is also sent to the National Snow and Ice Data Cen-ter (NSIDC), repackaged into additional user-friendly formats such as Keyhole Markup Language (KML) andESRI Shapefiles, and made available to the public as the Multisensor Analyzed Sea Ice Extent (MASIE) prod-uct [Fetterer et al., 2010].)

As discussed in Posey et al. [2015], the effective resolutions of observed sea ice concentration from SSMI/Sand AMSR2 are 25 km and 10 km, respectively. These observed values must be interpolated to the higherresolution model grid (Figure 1). With model horizontal resolution 3.5 – 4.0 km north of 708N, the relativecoarseness of these data can lead to shifts in observed ice edge location compared to the model by 2–6grid points, on the order of 10 km – 20 km. While the use of the IMS mask at 4 km resolution helps mitigatethe observed versus model resolution differences, IMS is a product that can vary depending on the

Figure 1. ACNFS horizontal grid spacing in km.

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 4

availability of satellite coverage on a particular day. For example, SAR or VIS/IR observations are not avail-able at all locations on all days. Also, IMS is a product created by human analysts, and subjective differencesin ice location can occur [Meier et al., 2015].

The 4 km resolution IMS is used to mask the ice concentration assimilated into ACNFS, as 4 km is similar tothe ACNFS horizontal resolution. IMS is a mask of ones and zeros that identify where the ice concentrationis greater than 40%; there is no ice concentration value included within IMS. Since ACNFS requires anumeric ice concentration value for data assimilation, a method is required to convert the IMS mask to anice concentration. Such a method was developed and described in Posey et al. [2015], where the IMS maskis applied to the NCODA analysis ice concentration, A, as follows:

A5

0%;

70%;

A;

IMS50

IMS51 and A < 70%

IMS51 and A � 70%

8>>>><>>>>:

(1)

In (1), if IMS 5 0, the NCODA analysis ice concentration is removed and made 0%. If IMS 5 1 and NCODAanalysis ice concentration is less than 70%, NCODA analysis ice concentration is made to be 70%. IfIMS 5 1 and the NCODA analysis ice concentration is greater or equal to 70%, the NCODA analysis iceconcentration value is not changed. The threshold of 70% was chosen as the midpoint of the range ofvalid IMS ice concentration values (40%–100%) [Posey et al., 2015]. This blending algorithm in (1) was cre-ated in conjunction with the NSIDC [Fetterer et al., 2015]. The difference is that Fetterer et al. [2015] apply(1) to AMSR2 ice concentration, whereas in ACNFS equation (1) is applied to the NCODA analysis ice con-centration. The blending method of IMS and NCODA analysis ice concentration was first introduced toACNFS 2 February 2015. This date is noted in the forecast assessments to investigate the effect of IMS onACNFS forecasts.

2.5.2. Assimilating Sea Ice ConcentrationOnce the sea ice analysis field is created (blended with AMSR2 and IMS after February 2015), it is assimilatedinto ACNFS using the following algorithm:

1. If model<NCODA analysisa. Use NCODA analysis for model< 15%.b. Linearly blend model and NCODA analysis where (15%<model< 30%)c. Use model where model> 30%

2. If model�NCODA analysis

a. If NCODA< 1%, remove all ice.b. Otherwise:

i. Use NCODA analysis where NCODA analysis< 25%ii. Linearly blend model and NCODA analysis where (25%<NCODA< 50%)

iii. Use model where NCODA analysis> 50%

This assimilation method varies depending if the initial model ice concentration is less than or greater thanthe NCODA ice concentration. The algorithm is designed to assimilate near the ice edge, the area of mostinterest for U.S. Navy and other maritime operations. The main reasons for choosing this algorithm are two-fold: first, it has been determined that sea ice obtained from SSMI/S using the Navy Cal-Val algorithm over-estimates ice concentration in the central Arctic [Hollinger, 1991]; second, passive microwave sensors suchas SSMI/S and AMSR2 tend to underestimate ice concentration in regions of thin ice or where meltpondsexist on the sea ice surface [Meier et al., 2015]. The result is a weighted average of the NCODA sea ice analy-sis field and the modeled initial conditions. As seen in the algorithm, the threshold where ice is assimilatedvaries depending on whether the sea ice analysis field is greater or less than the model initial condition. Incase 1 of the algorithm, the model ice concentration is less than NCODA, and thus ice needs to be added tothe model. We only add ice in the vicinity of the modeled ice edge, where the model ice concentration isless than 30%. Where the model is less than 15%, NCODA is directly inserted into the model. Where themodel is between 15% and 30%, a linear weighting is applied that puts more weight on the model valuethe closer it is to 30%. Above 30% the model value is retained. In case 2 of the algorithm, the modeled iceconcentration is greater than NCODA, thus ice needs to be removed. Where the NCODA ice concentration

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 5

is less than 25%, it is directly inserted into the model. WhereNCODA is between 25% and 50%, a linear weighting isapplied between the model and NCODA with more weighton the model as the NCODA value increases. Where NCODAis greater than 50% the model value is retained. In addition tothe change in ice concentration, the modeled sea surfacetemperature is adjusted to be slightly above or below thefreezing point, based on the addition or removal of ice, toprevent the model from immediately melting or forming seaice based on sea surface temperature.

In regard to assimilating NCODA blended with IMS, we notethat after blending the minimum NCODA ice concentrationvalue (other than zero) is 70%. In the instance where themodel ice concentration is less than 15% and IMS51(NCODA 5 70% ice), a value of 70% is inserted into the model.Nearby there may be a model value of 30%, where theNCODA value is not assimilated and the 30% is retained. Thiscould lead to regions with sharp gradients between 70% and30%. In practice, this situation occurs rarely, as the regionwhere ice changes from 15% to 30% is usually very narrow.Also, since this method is focused near the ice edge wherethe change is greatest, daily assimilation updates in the iceedge minimizes the long-term impact.

2.6. Model SetupACNFS has a horizontal resolution of 3.5–4.0 km in the polarlatitudes north of 708N (Figure 1). South of 708N, the horizon-tal resolution gradually increases to 7.5 km at 408N. At theopen boundary at 408N, in the Atlantic and Pacific Oceans,Neumann (reflective) boundary conditions are applied inCICE, while in HYCOM the boundary conditions are obtainedfrom a real-time 1/128 global HYCOM/NCODA forecast systemwith an embedded thermodynamic sea ice model in place of

CICE [Metzger et al., 2008]. Neumann boundaries are used in CICE since 408N is adequately far away fromany sea ice covered region. HYCOM model bathymetry is based on the NRL 2 min Digital Bathymetric Data-base (DBDB2, http://www7320.nrlssc.navy.mil/DBDB2_www/).

Figure 2 contains a flowchart of how ACNFS is run each forecast period. For each 7 day forecast period, a 3day hindcast is performed where ACNFS is first started 3 days prior to the current date. During this 3 dayhindcast is when observational data are assimilated into the system. The reason for this is two-fold: (1) toassimilate data that may only be available 2 or 3 days after the observation time, and (2) to minimize theshock of replacing data in the model by starting back in time and gradually running to the assimilatedvalue. After this 3 day hindcast, data are no longer available for assimilation, and a 7 day forecast is run.Note that the hindcasts are started at 18Z, while the forecasts are generated at 00Z. Thus, the first outputfrom an 18Z start time is at 00Z the next day, making it a 6 h forecast. For the remainder of this paper fore-cast times will be referred to by hour (6 h, 30 h, etc.) instead of day (1 day, 2 days, etc.).

3. Forecast Assessments

3.1. Ice Concentration Forecast SkillIn assessing the ice concentration forecast skill of ACNFS, a skill score using the mean square error is com-puted based on the analysis of Murphy and Epstein [1989] and Van Woert et al. [2004]. Following the nota-tion of Van Woert et al. [2004], skill score is generically defined as:

H3

NCODA OBS

H2

NCODA OBS

H1

NCODA OBS

F1

F2

F3

(18Z)

(18Z)

(18Z)

(00Z)

(00Z)

(00Z)

(00Z)

Figure 2. Flowchart of the hindcast/forecast process,where H5Hindcast, F5Forecast. The number indi-cates day of hindcast or forecast, so H3 5 3 dayHindcast, F1 5 1 day Forecast, etc. The number inparenthesis indicates the model hour. Thus, startingfrom H1, the first forecast time at 00Z is a 6 h fore-cast. The next forecast is the next day at 00Z, makingit a 30 h forecast, etc.

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 6

SS5kf 2kR

kp2kR(2)

where kf , kp, kR are the accuracy of the forecast, the accuracy of a ‘‘perfect’’ forecast, and the accuracy of areference forecast, respectively. Accuracy is defined in terms of the mean-square error (MSE) [Murphy andEpstein, 1989], defined here as:

MSE a; bð Þ5 1N

XN

i51

ai2bið Þ2 (3)

where ai , bi are values at location i, and N is the number of locations examined. (Selection of the i locations isdescribed below.) Substitution of (3) into (2), and noting that MSE of a perfect forecast is zero (since a5b), yields:

SS512MSE f ;Oð ÞMSE R;Oð Þ (4)

where f is the model forecast, O is an observed ice concentration, and R is a reference ice concentrationused for comparison with the forecast. In (4), when the forecast MSE is less than the reference MSE, SS ispositive and approaches 1. In contrast, if the forecast MSE is equal to or greater than the reference MSE, SSis zero or negative. Thus, SS> 0 indicates ‘‘skill’’ or improvement compared to reference field, whereasSS� 0 indicates no skill. In this paper two reference states are examined: climatology and persistence.

It is important to note that in (4), the observed field, O, in this study is the assimilated analysis field after theapplication of the assimilation algorithm described in section 2.5.2, and not the raw satellite ice concentra-tion nor the raw NCODA analysis field. The reason for the use of the assimilated ice concentration for O isthat the assimilation technique is designed to be our best estimate of the ‘‘true’’ ice coverage, particularlynear the ice edge, given known biases in satellite ice concentration observations. Also, since the NCODAanalysis field is modified by the weighting technique, the model never starts with an ice concentration thatis the same as the observation field. Thus, using the raw observation would not be consistent with how themodel is run.

The skill score relative to climatology is a measure of the skill of the model ice concentration forecast com-pared to assuming a climatological value. From (4), the climatology skill score is defined as

SSc nð Þ512MSE fn; Anð ÞMSE Cn;Anð Þ (5)

where n is the model forecast hour n5(6, 30, 54, 78, 102, 126), fn is the corresponding ice concentrationforecast field, Cn is the ice concentration from the ECMWF ice concentration climatology, and An is theassimilated ice concentration obtained as in section 2.5.2. In (5), O is replaced by An seen in (4) to note theuse of the assimilated ice concentration.

The skill score relative to persistence is a measure of model forecast ice concentration compared to assum-ing the conditions at the initial time will persist throughout the forecast period (i.e., conditions will stay thesame). Here the skill score for persistence is defined as

SSp nð Þ512MSE fn;Anð ÞMSE A0;Anð Þ (6)

where A0 is the initial assimilated ice concentration (section 2.5.2) at the start of the forecast, and the otherterms are as previously defined above.

In selecting the locations of interest, i, in (3), we are interested in the skill of ACNFS where the ice concentra-tion changes. In this study, the locations were determined by identifying where the assimilated ice concen-tration changes by 65% or more over 5 days. This is similar to the selection of points performed in Lemieuxet al. [2015], and is simply the magnitude of the difference between the assimilated ice concentration at thestart of the forecast and the assimilated ice concentration from the run started 5 days later. This methodcaptures both ice formation and melt. Sample locations are shown in Figure 3 for March 2014 and Septem-ber 2014, where it is seen that the points are located along the ice edge and marginal ice zone. It is impor-tant to note that the assimilation scheme described in section 2.5.2 only assimilates ice concentration

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 7

Figure 3. Areas where the assimilated ice concentration changes by more than 65% for (top) 10–15 March 2014 and (bottom) 10–15September 2014.

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 8

where the model ice concentration is less than 50%. Thus, in both plots of Figure 3, there are no locationsof interest in the central ice pack since there was no assimilation of sea ice. In addition to limiting the num-ber of points to those that change over the forecast, another advantage of this method is that the numberof points used for each forecast on a particular day is the same (the number of points changes from day today, however). Also, as noted in Lemieux et al. [2015], the CICE model is run at very high spatial resolutionand produces linear kinematic features (LKFs) that are not seen in the lower resolution satellite observa-tions. If the raw satellite data were to be used to locate points of interest, changes in the modeled LKFswould be missed. Thus, using the assimilated data to locate model points of change is more appropriatethan the raw satellite observations. Since the points are based solely on the change in analysis field, withoutany consideration of the change in forecast field, points where the forecast changes by 65% or more arenot taken into account. As a result, our analysis does not take into account ‘‘false alarms’’ when the modelforms ice where no ice forms in the analysis fields [Van Woert et al., 2004]. Daily skill scores were computedfor the time period February 2014 to June 2015. The number of data points (not shown) varied from approx-imately 85,000 in the middle of September, during the minimum ice extent, to near 300,000 in May, whenthere is maximum ice area available to melt.

Figure 4 contains plots of the monthly mean SSp and SSc from the daily skill scores. In each plot there are twohorizontal dashed lines at 0 and 0.36. Values above zero indicate model skill compared to the reference field,while values above 0.36 are considered to be particularly skilful [Van Woert et al., 2004]. The value 0.36 is chosenas corresponding to the selection of 0.6 correlation coefficient value as skilful in Hollingsworth et al. [1980]. InFigure 4a, SSp is shown to always be above zero, indicating skill compared to assuming a persistent ice state.Also, all forecasts except the 6 h forecast are above 0.36, indicating good skill compared to persistence. The SSp

value below 0.36 for the 6 h forecast (prior to assimilation of AMSR21IMS) is attributed to the relatively slowevolution of the ice concentration field, which typically does not change significantly in a 6 h period.

SSc is also seen to be above zero in Figure 4b, with the exception of the 126 h forecast between September2014 and January 2015. As one would expect, the longer the forecast the lower the skill score compared toclimatology, as errors associated with the forecast increase. An interesting feature of this plot is that, basedon this short 16 month analysis, SSc tends to be cyclic, with higher scores in the summer months (July–September) up to the summer minimum ice extent and lower scores during the winter freeze up period of

Figure 4. Skill scores relative to (a) persistence and (b) climatology. Scores above 0.0 show skill, scores above 0.36 are considered particularly skilful. The vertical dashed line is whenACNFS started assimilating ice concentration masked by IMS.

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 9

October–January. The higher SSc values in the summer suggest ACNFS does a good job at capturing icemelt, while degradation in SSc during the freeze up period suggests that ACNFS may not form ice as fast asactually occurs in the Arctic.

In (5), the low SSc in the winter can be attributed to either an increase in the forecast error or a decrease inthe climatology error. Figure 5a depicts the bias between the climatological ice concentration and theassimilated value for that date, Cn2Anð Þ, while Figure 5b is a plot of the bias between the model forecastand the assimilated value for that date, fn2Anð Þ. Here biases are examined instead of MSE to investigate thesign of the differences, since MSE in (3) is by definition a positive number. In Figure 5a, from May to January,the climatological bias is positive, indicating the climatology overestimates the ice concentration. This is inagreement with the observed decline in September sea ice minimum compared to climatology. It is inter-esting to note that in the melt period, May–September, the climatology bias increases with each forecasttime. The increase in bias with forecast time occurs since, during the summer months, the ice is meltingfaster than seen in climatology. The converse is seen as the freeze season starts in October–January; the cli-matology bias decreases with forecast time as the ice grows and becomes closer to the climatology concen-tration. In contrast, the forecast bias (Figure 5b) during the winter freeze months of September–Decemberis negative, indicating the model has less ice than the reference observation. The forecast bias becomesmore negative with forecast time, indicating the model does not freeze ice as fast as observed. This, alongwith the reduction in climatology bias at each forecast time during the winter months, explains the degra-dation in SSc during this time. Several possibilities exist to explain the large negative bias in forecast bias,ranging from perhaps too much incoming atmospheric radiation to warmer ocean temperatures to sea iceparameterizations. These will be the topic of a future study. We remind the reader that these biases arecomputed over the points where ice changes by 65% over a 5 day period, and not the entire Arctic.

The assimilation of ice concentration with the blended AMSR21IMS products was initiated in ACNFS on 2February 2015. This date is marked as a vertical dashed line in Figures 4 and 5. The skill score versus persist-ence (SSp) increased significantly for the 6 h forecasts, while SSp remained virtually unchanged for the 30 hforecasts and dropped slightly for forecasts past 54 h. Since skill scores are a ratio of MSE’s, the drop in theskill scores after 54 h forecast is caused here by a greater reduction in the persistence MSE compared to the

Figure 5. Monthly mean ice concentration (%) bias compared to the assimilated ice concentration for (a) climatology and (b) each forecast. The climatology bias is large in the wintermonths, indicating climatology has more ice than observations. The forecast bias is negative in the winter, indicating the model has less ice than observations. The vertical dashed line iswhen ACNFS first started to assimilate ice concentration masked with IMS.

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 10

reduction in the forecast MSE (not shown). It is noted that the results with IMS are preliminary, as we haveonly been assimilating with the blended AMSR21IMS for 4 months at the time of this writing. A longer-term study is needed to make more definitive conclusions regarding the impact of IMS assimilation.

3.2. Ice Edge LocationThe NIC Ice Edge product is a product intended for navigational purposes to help vessels avoid nearly all icehazards, and is created independently of the IMS product [Helfrich et al., 2007; Posey et al., 2015]. The NICIce Edge product is a series of latitude/longitude points defined where the ice concentration ‘‘variesbetween 0 and 1/10th’’ (S. Helfrich, NIC, personal communication 2014), and is a more conservative defini-tion of ice edge compared to IMS ice extent The NIC Ice Edge product locations are determined by an ana-lyst and typically utilizes different data sources, including satellite and in-situ observations, to determine theice edge location compared to IMS. Figure 6 contains a representative plot of the difference between IMSand the NIC Ice Edge product. Here it is seen that the NIC Ice Edge product locations are almost always fur-ther south and contains more ice area than IMS. On this particular date shown in Figure 6, a polynya is seenin the Beaufort/Chukchi seas and is identified by IMS, whereas the NIC Ice Edge product encompasses theentire polynya region. Since the NIC Ice Edge product locations are not assimilated into ACNFS, we compare

Figure 6. Difference between IMS and NIC Ice Edge Locations on 25 August 2015. The colored area is 10 km gridded ice concentration (%) from AMSR2. The black line is the IMS mask,and the purple dots are the NIC Ice Edge. The right plots contain the same data as the left, but are zoomed in to the Beaufort/Chuckchi Seas and the Fram Strait. The more conservativeNIC Ice Edge location is almost always further south compared to IMS.

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 11

the distance between the NIC Ice Edge product locations and the ice edge location derived from ACNFS.(See Posey et al. [2015] for a more detailed discussion of the differences between the NIC Ice Edge productand IMS ice extent.)

In this study the ACNFS ice edge is defined as those points that exceed a specified threshold value for iceconcentration that also have a neighboring point that falls below this threshold value. The midpoint of theNIC Ice Edge product ice concentration definition, 5%, is used as the ACNFS ice concentration threshold.Thus, any model grid cell with ice concentration greater than 5% that also has an adjacent cell with ice con-centration less than 5% is identified as the model ice edge. After determining the ACNFS ice edge, the nextstep is to identify the ACNFS model grid point that contains the NIC Ice Edge product location. The closestACNFS ice edge point to the NIC Ice Edge product location is found, and the distance between the twopoints is then computed using via the Haversine formula [Gade, 2010, equation (19)]

De52r arcsin

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffisin2 U12U0

2

� �1cos U0ð Þcos U1ð Þsin2 H12H0

2

� �s !(7)

Here r is the radius of curvature of the Earth, U and H are latitude and longitude (respectively) in radians, andthe subscripts 0, 1 indicate either ACNFS ice edge or NIC Ice Edge, respectively. Assuming the Earth to be anellipsoid, the radius of curvature is dependent on latitude and computed as in Snyder [1987, equations (4)–(18)]:

r5a 12e2ð Þ

12e2 sin2 Uð Þ (8)

where a56378137.0 m is the semi-major axis length of the Earth, e58.181919 3 1022 is the Earth’s eccen-tricity obtained from WGS84 standard [NIMA, 2000], and U is the average latitude between the two ice edgepoints. The value of De in (7) is an absolute value of the distance between NIC Ice Edge product locationand ACNFS ice edge is used for analysis. Thus, there is no signed value to indicate if the ACNFS ice edgeextent is more or less than the NIC Ice Edge product location. Since there may be a spatial dependence onthe ice edge difference, daily mean ice edge distances for seven regional seas are computed: Central Arctic,Greenland-Iceland-Norwegian (GIN) Seas, Barents Sea, Laptev Sea, Sea of Okhotsk, Chukchi Sea, and theCanadian Archipelago (Figure 7a).

The ice edge error results for the entire time of this study, February 2014 to June 2015, are shown in Figure7b. The forecast ice edge distance error is shown in solid lines, while the ice edge distance error from per-sistence is shown in the dashed lines. Here it is seen that persistence has similar or lower ice edge errorscompared to the forecasts. The lower error by assuming persistence indicates the modeled ice edge inACNFS evolves more rapidly than in the Arctic, and the modeled error is greater than assuming a persistentice state (this changes with the assimilation of IMS, discussed below). Overall, the ice edge errors rangefrom near 30 km at 6 h forecast to over 80 km for 126 h forecast. As expected, the ice edge error distanceincreases as forecast time increases. Depending on the region, the forecast edge error increases by 10 km(Barents Sea) to 20 km (Canadian Archipelago) over the forecast period. The Canadian Archipelago, with thepresence of many small straits and landmasses, is the most challenging of regions to model, and thus it isexpected this region will have the highest ice edge error.

To look at the effect of assimilating the blended AMSR21IMS on the ice edge error, Figure 7c is a plot of iceedge error for the months February–June 2014, before ACNFS started to assimilate the blended product,while Figure 7d is a plot of ice edge error for the same months when ACNFS assimilated the blendedAMSR21IMS product, February–June 2015, for comparison. In Figure 7d, the influence of assimilating theblended AMSR21IMS can be seen, the forecast ice edge is closer to the NIC Ice Edge and errors usingAMSR21IMS are lower than persistence for all regions except the Canadian Archipelago 102 h and 126 hforecasts. It is interesting that in Figure 7d, the forecast ice edge error (solid lines) decreases through thefirst 54 h. This decrease can be explained by noting that the months of the study with AMSR21IMS (Febru-ary–June), the ice is retreating. Since the IMS is not as conservative in identifying the ice edge as the NIC IceEdge (Figure 6), the ACNFS initial extent using AMSR21IMS is less than the NIC Ice Edge. As time evolves,the daily NIC Ice Edge location would be retreating faster than the model, causing the ice edge comparisonto improve with time. Although this pattern is encouraging, it is also noted that in some regions the initialmagnitude of the error increased by 5–10 km compared to not assimilating AMSR21IMS, including the GIN

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 12

Seas and Barents Sea (although the initial error value increased, the actual change in error over the 126 hforecast was roughly the same for each region). This increase in ice edge error is on the order of 2–5 ACNFSgrid points and could be a result of inter-year model variability or NIC Ice Edge location variability (recall theNIC Ice Edge is generated via human analyst, and analysis by person can vary) [Posey et al., 2015].

3.3. Ice DriftIn addition to forecasting the ice edge location, it is of interest to forecast ice motion. This is important for pre-dicting where the ice is heading, and is particularly useful for navigation planning purposes. Here we comparethe ACNFS ice velocity and separation distance to data obtained from the International Arctic Buoy Pro-gramme (IABP). IABP is a network of drifting buoys in the Arctic Ocean to provide meteorological and oceano-graphic data for real-time operational requirements and research purposes; data are made available to thepublic at http://iabp.apl.washington.edu. For the time period of this study, 377 buoys were used. Two quanti-ties are investigated: (1) ice drift speed and (2) separation distance error. For each quantity the forecast differ-ences are compared to values obtained if persistence was assumed. Each is described further below.3.3.1. Drift speedThe IABP speed is determined by the distance the buoy traveled over 24 h. The distance the buoy traveled,Db, is obtained via the Haversine formula defined in (7), but with a slight modification of the latitude andlongitude subscripts:

Figure 7. (a) Regions defined for the ice edge analysis. (b) Average ice edge forecast error (km) compared to the NIC ice edge for each region. (c) Same as Figure 7b but from Februaryto June 2014 (without IMS). (d) Same as Figure 7c but from February to June 2015 (with IMS). Solid lines represent forecast error, dashed lines represent persistence error. Note thechange in vertical axis from Figure 7b to Figures 7c and 7d.

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 13

Db52r arcsin

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffisin2 U242U0

2

� �1cos U0ð Þcos U24ð Þsin2 H242H0

2

� �s !(9)

Here the subscripts 0, 24 indicate the buoy starting position time and position time 24 h later, respectively,and all other terms are defined as in (7) and (8). The speed of the buoy in meters/second is then Db dividedby 86,400 seconds per day:

Sb5Db

86400(10)

Here we note that Sb is a mean speed over 24 h. In contrast, the velocity fields written by ACNFS are instantane-ous, and not a daily average. In order to define a modeled speed that is similar to a daily mean, an average ofthe modeled output velocity fields, u and v, at the start of the analysis time and that 24 h later was computed:

Sm5ffiffiffiffiffiffiffiffiffiffiffiffiffiffi�u21�v 2

p(11)

Here the overbar represents the average of the current velocity and the velocity 24 h later. It is important tonote that, due to the fact that ice velocity is highly dependent on the atmospheric stress applied, and thatstress is provide from NAVGEM every 3 h, there is an error associated with computing the average onlyfrom instantaneous outputs 24 h apart.

In order to compare the modeled and IABP speeds, the first step is to obtain the model grid cell containingthe buoy location at the particular model output date. The difference between the model forecast ice speed

Figure 8. Monthly mean ice drift speed difference (cm/s) between IABP and ACNFS. Blue line is persistence difference, red line is forecastdifference. The monthly means were computed from all available IABP drifters in that month. The model restart ice velocity is used as thereference state for persistence. The forecast time (tau) is noted in the title of each plot. The location of the IABP was linearly interpolatedto the ACNFS grid to obtain the ACNFS ice drift speed.

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 14

and buoy ice speed, Df 5Sfm2Sf

b, is then computed for each forecast time (here f is the forecast hour). Inaddition, the speed difference is computed by assuming a persistent ice state from the forecast initializa-tion, Dp5S0

m2Sfb, where the superscript ‘‘0’’ indicates the forecast initialization. Figure 8 contains the

monthly mean Df (red line) and Dp (blue line) for each forecast time. For all the forecast times the model isslower than the buoy speed in the summer and autumn months (June–October), with a max velocity differ-ence around 22 cm/s in August. In contrast, the winter and spring months (November–May) the model isfaster than the buoys, except for April past 54 h. It is also interesting to note that during summer andautumn, persistence performs better than the forecast for forecast times at and beyond 54 h, while for thewinter and spring months the opposite is generally the case. Since ice drift is largely driven by wind andocean stresses, these results suggest that there may be seasonality biases in the wind forcing or air-ice/ocean-ice drag coefficients. This will be the topic of a future study.3.3.2. Drifter separationThe drifter separation difference is the distance between the actual drifter position and the projected posi-tion using the ACNFS velocity field for each forecast period. Since the distance traveled is dependent on thelength of forecast time, a comparison with equal spaced times is desired. Also, since the IABP buoy locationsare reported twice per day at 00Z and 12Z, it is convenient to start the analysis with the 6 h forecast insteadof the initial velocity field. This avoids the need to interpolate the IABP position to the model restart time(18Z). Therefore, the drifter location comparisons were started with the 6 h forecast and made with everysubsequent 24 h forecast. This was accomplished by first identifying the location of the buoy at the 6 h fore-cast time. The ACNFS forecast velocity at this location was obtained by making a triangular interpolant of

Figure 9. Monthly mean 24 h drifter separation distance (km) between forecast (red) and persistence (blue) buoy positions starting ateach forecast tau. The forecast separation analysis starts with the 6 h forecast time since IABP data are available at 00Z. As a result, the 6 hforecast velocity is used for persistence. Starting with 6 h forecast, the location of the IABP buoy is found in ACNFS, then the ACNFS veloc-ity (forecast or persistence) is applied for 24 h. The difference between the modeled position and actual IABP position 24 h later is thencomputed. Then at the 30 h forecast, the modeled ice drift is applied for 24 h to the resulting position from the 6 h forecast. The new resultis compared to the IABP position 24 h later. The process is continued for all forecast times.

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 15

the ACNFS grid, then applying a ‘‘natural neighbor’’ interpolation scheme [Ledoux and Gold, 2005]. The pro-jected position of the virtual buoy is then ‘‘forecast’’ using the ACNFS velocity at this time for the next 24 h,and the distance between this forecast location and the actual buoy location is computed. The process isrepeated for the next 24 h period, using the velocity from the forecast position (i.e., not the actual buoylocation) to update the forecast buoy position. Note that as defined in (11), the forecast velocity applied ateach period is an average of the current velocity and the velocity 24 h later, not the instantaneous velocityfield. In addition to the forecast distance, the position difference was computed by assuming a persistentvelocity using the 6 h forecast field for the entire 126 h forecast period.

Figure 9 contains the monthly averaged distances between observed and forecast buoy location. For the 6 hforecast, the buoy distance is about 6 km for both persistence and forecast velocities (recall that the 6 h fore-cast is the reference time, thus the results are the same for both forecast and persistence). As forecast timeincreases, the distance between modeled and actual buoy location increases. The forecast error is lower thanwhen assuming persistence, reaching an average of 30 km by the 126 h forecast compared to an average dis-tance error of 45 km by assuming persistence. This is in contrast to the ice velocity differences shown in Figure8, where persistence out-performed forecast velocity, especially in the summer months. Even though themodel was slower than persistence, the better performance by the forecast in the drifter distance analysis isdue to the inclusion of the direction of the ice drift in each forecast. Since the forecast drift adapts with timecompared to persistence, the modeled forecast location is closer to the actual buoy location than assuming apersistent ice state. Similar to the speed difference above, the largest distance error seems to be in thesummer and autumn months. Factors that influence the forecast separation are the atmospheric forcing anddrag coefficients, as well as potential biases in ice seasonal ice thickness and ice strength that could lead tobias in drift speed [Zhang et al., 2012], and will be examined with a longer-term study.

4. Summary

Forecast skill can broadly be thought of as any improvement the model provides relative to assuming somebaseline condition. Two baseline conditions are considered in this study: persistence (assuming conditionsdo not change over the forecast time), and climatology (historical averages). In determining ACNFS forecastskill, we are interested mainly in the marginal ice zone where the ice concentration changes the fastest withtime. Therefore, forecast skill was computed only where the assimilated ice concentration changed by morethan 65% over a 5 day period. This captures areas of ice growth and melt, and is limited mostly to the mar-ginal ice zone (Figure 3). Starting in February 2015, ACNFS began to assimilate a blended AMSR21IMS prod-uct. The effect on skill scores with this new data set is also examined.

Overall, ACNFS demonstrates a high level of skill compared to persistence, with SSp values above zero forthe entire time of this study (Figure 4a). Only the 6 h SSp is below 0.36. Over a 6 h period the ice conditionsare not expected to change much, thus the skill at 6 h compared to persistence is expected to be lower.One would expect SSp to increase as the forecast time increases, but that is not the case here; at the 30 hforecast point and beyond SSp hovers around 0.4 2 0.5. While this still indicates the forecasts are betterthan assuming a persistent ice state, further investigation as to why the SSp does not increase past 30 h isneeded. The inclusion of IMS to the assimilated NCODA ice concentration improved the 6 h SSp to about0.4, making it ‘‘particularly skilful.’’ Unfortunately, the inclusion of IMS seemed to have little effect on theother forecast times. Since this was only performed for a 4 month period, a longer study is needed to evalu-ate any potential biases throughout the year.

The skill score compared to climatology, SSc , behaves as one would expect, with SSc the highest for the 6 hforecast and decreases with forecast time (Figure 4b). SSc is positive for all forecast times except for the win-ter months of the 126 h forecast, and thus performs better than assuming a climatological ice concentra-tion. SSc can be affected by inter-annual variation in sea ice cover. If the sea ice concentration of a particularyear happens to be close to climatology, then SSc will be lower, since MSE Cn;Anð Þ in (5) will decrease. Con-versely, SSc will increase if the ice concentration departs greatly from climatology. In order to examine thedip in SSc from October to January, the climatology and forecast biases are examined (Figure 5). In Figure5a, the climatology is shown to be biased higher than the assimilated ice concentration during May–January, as one would expect given the increased rate of summer sea ice extent decline. The climatologybias decreases with forecast time from October to January (Figure 5a), as the model starts to form more ice

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 16

and become close to the climatology value. The bias in forecast ice concentration increases with forecasttime (Figure 5b) during the same time, indicating that model does not form ice as quickly as observed. Sev-eral possibilities exist to explain the large negative bias in forecast bias, ranging from perhaps too muchincoming atmospheric radiation to warmer ocean temperatures to sea ice parameterizations. Since the skillscore is the ratio of the MSE of these terms, the combination of these facts is causing the skill score todegrade during this time for the longer forecasts. Also, by defining the points of interest to be where theassimilated ice concentration differs by greater than 65% over a 5 day period, the comparison is limited tonear the ice edge and does not indicate model skill compared to climatology for the full Arctic.

Regarding SSc , the sea ice climatology used in this analysis was created from ECMWF data for the years1979–1996. As noted in the introduction, the past few decades have seen a greater reduction in Septembermean ice extent compared to earlier decades. While the use of this climatology was consistently used toquality control the ice observations, it is possible that using a more recent climatology from either ECMWFor the National Centers for Environmental Prediction (NCEP) could have an impact on SSc .

The location of the sea ice edge is important for operational planning of science and maritime activity. Inthis study the forecast ACNFS ice edge (defined as 5% ice concentration) is compared to the independentlyobtained NIC Ice Edge product locations. The analysis was divided into 7 regions (Figure 7a). ACNFS per-formed best in the Barents Sea region, with ice edge error ranging from 30 km at 6 h forecast to 40 km at126 h forecast, and worst in the Canadian Archipelago, with ice edge errors ranging from 60 km at 6 h fore-cast to 80 km at 126 h forecast. With many small openings and landmasses in the Canadian Archipelago,this region is perhaps the most difficult to model. In order to examine the effect of assimilating IMS, Figures7c and 7d contain plots of the ice edge error for the same length of time (February–June) before and afterassimilating AMSR21IMS. Here it is seen that the inclusion of AMSR21IMS reduced the ice edge error com-pared to the non-IMS period. In Figure 7d the error is also seen to decrease with time up to the 54 h fore-cast. This is a consequence of the IMS field not defined as conservative as the NIC Ice Edge locations (Figure6). During this spring ice retreat, as forecast time lengthens the NIC Ice Edge locations will move closer tothe IMS location, causing the ice edge error stats to improve.

In addition to ice concentration, ice speed and buoy separation distance were compared to data reportedfrom IABP buoys. Overall, the ice velocity in ACNFS was slower than IABP buoys in the summer months ofJune–October (Figure 8). During this time, assuming a persistent ice speed performed better by as much as1.5 cm/s for the September 126 h forecast. In contrast, the modeled velocity was greater than observed forthe March–May time frame. This suggests a seasonal bias to either the wind or drag coefficients. This couldoccur as a result of the change in surface conditions. For instance, in the summer months most of the snowhas melted and could result in a rougher surface than when snow exists and fills surface ridges and cracks.This change would not be taken into account in ACNFS, and the wind stress could be underestimated.

In addition to ice velocity, a virtual IABP buoy was modeled by starting at the same position as the real IABPbuoy, but advected using the ACNFS forecast ice drift velocities. In contrast to the velocity difference, themodeled drifter separation distance was smaller for each forecast compared to persistence. This impliesthat while the ACNFS ice speed may be slower or faster during a particular time, the forecast ice direction iscloser to reality than if a persistent ice speed and direction are assumed.

ReferencesArrigo, K., G. van Dijken, and S. Pabi (2008), Impact of a shrinking Arctic ice cover on marine primary production, Geophys. Res. Lett., 35,

L19603, doi:10.1029/2008GL035028.Aschan, M., and R. Ingvaldsen (2009), Recruitment of shrimp (Pandalus borealis) in the Barents Sea related to spawning stock and environ-

ment, Deep Sea Res., Part II, 56(21-22), 2012–2022, doi:10.1016/j.dsr2.2008.11.012.Bertino, L., and K. A. Lisæter (2008), The TOPAZ monitoring and prediction system for the Atlantic and Arctic Oceans, J. Oper. Oceanogr.,

11(2), 15–18.Bitz, C. M., and W. H. Lipscomb (1999), An energy-conserving thermodynamic model of sea ice, J. Geophys. Res., 104(C7), 15,669-15,677, doi:

10.1029/1999JC900100.Bleck, R. (2002), An oceanic general circulation model framed in hybrid isopycnic-Cartesian coordinates, Ocean Modell., 4(2), 219–219, doi:

10.1016/S1463-5003(01)00017-8.Buehner, M., A. Caya, L. Pogson, T. Carrieres, and P. Pestieau (2013), A new environment Canada regional ice analysis system, Atmos. Ocean,

51(1), 18–34, doi:10.1080/07055900.2012.747171.Campbell, T., et al. (2010), Integrated modeling of the battlespace environment, IEEE Comput. Sci. Eng., 12(5), 36–45, doi:10.1109/MCSE.2010.78.Carnes, M., R. W. Helber, C. N. Barron, and J. M. Dastague (2010), Validation test report for GDEM4. Tech. Rep. NRL/MR/7330–10-9271, Naval

Res. Lab., Stennis Space Center, Hancock, Miss.

AcknowledgmentsIABP data are made available to thepublic at http://iabp.apl.washington.edu. AMSR2 data are available afterregistration at GCOM-W1 data serversite https://gcom-w1.jaxa.jp. IMS andNIC ice edge products are availablefrom the NIC web site http://www.natice.noaa.gov/Main_Products.htm.ECMWF sea ice concentration data areavailable via http://www.ecmwf.int/en/research. This work was funded as partof the Naval Research Laboratory’s 6.1Research Option ‘‘Determining theImpact of Sea ice Thickness on theArctic’s Naturally ChangingEnvironment’’ (DISTANCE). Numericalsimulations were performed at theNavy Department of DefenseSupercomputing Resource Center(DSRC) using grants from theDepartment of Defense HighPerformance ComputingModernization Program. All modeldata are securely stored at the NavyDSRC archive server. The files storedthere can be accessed after obtainingan account at the facility. Thecorresponding author can becontacted for information to accessthe archived data once an account hasbeen established. The authors thankAxel Schweiger and Jean-FrancoisLemieux for their comments leading toseveral improvements in this paper.

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 17

Cayula, J-F. P., D. A. May, B. D. McKenzie, K. D. Willis (2013), VIIRS-derived SST at the Naval Oceanographic Office: From evaluation to opera-tion, in W. W. Hou and R. A. Arnone, Proceedings of SPIE 8724 on Ocean Sensing and Monitoring V, vol. 8724, 87240S–87240S-8, SPIE,Baltimore, Md., doi:10.1117/12.2017965.

Chassignet, E., L. Smith, G. Halliwell, and R. Bleck (2003), North Atlantic Simulations with the Hybrid Coordinate Ocean Model (HYCOM):Impact of the vertical coordinate choice, reference pressure, and thermobaricity, J. Phys. Oceanogr., 33(12), 2504–2526, doi:10.1175/1520-0485(2003)033< 2504:NASWTH>2.0.CO;2.

Comiso, J. C. (1986), Characteristics of winter sea ice from satellite multispectral microwave observations, J. Geophys. Res., 91(C1), 975–994.Comiso, J. C. and F. Nishio (2008), Trends in sea ice cover using enhanced and compatible AMSR-E, SSM/I, and SSMR data. J. Geophys. Res.,

113, C02S07, doi:10.1029/2007JC004257.Cox, M. (1984), A primitive equation, 3-dimensional model of the ocean, GFDL Ocean Group technical report 1, Geophys. Fluid Dyn. Lab.,

Princeton, N. J.Cummings, J. (2011), Ocean data quality control, in Operational oceanography in the 21st Century., edited by A. Shiller and G. Brassington,

pp. 91–122, Springer, Dordrecht, Netherlands.Cummings, J. A. (2005), Operational multivariate ocean data assimilation, Q. J. R. Meteorol. Soc., 131(613), 3583–3604, doi:10.1256/qj.05.105.Cummings, J. A., and O. M. Smedstad (2013), Variational data assimilation for the global ocean, in Data Assimilation for Atmospheric,

Oceanic and Hydrologic Applications, vol. II, edited by S. K. Park and L. Xu, pp. 303–343, Springer, Berlin, doi:10.1007/978-3-642-35088-7_13.

Durner, G., J. Whiteman, H. Harlow, S. Amstrup, E. Regehr, and M. Ben-David (2011), Consequences of long-distance swimming and travel overdeep-water pack ice for a female polar bear during a year of extreme sea ice retreat, Polar Biol., 34(7), 975–984, doi:10.1007/s00300-010-0953-2.

Fenty, I., and P. Heimbach (2013), Coupled sea ice-ocean-state estimation in the Labrador Sea and Baffin Bay, J. Phys. Oceanogr., 43(5),884–904, doi:10.1175/JPO-D-12-065.1.

Fernandez, P., G. Kelly, and R. Saunders (1998), Use of SSM/I ice concentration data in the ECMWF SST analysis, Meteorol. Appl. 5(4),287–296, doi:10.1017/S1350482798000942.

Fetterer, F., M. Savoie, S. Helfrich, and P. Clemente-Col�on (2010), Multisensor Analyzed Sea Ice Extent-Northern Hemisphere (MAISE-NH),National Snow and Ice data Center, Boulder, Colo., doi:10.7265/N5GT5K3K.

Fetterer, F., J. S. Stewart, and W. N. Meier (2015), MASAM2: Daily 4-km Arctic Sea Ice Concentration, 2012-2014, Natl. Snow and Ice Data Cent.,Boulder, Colo., doi:10.7265/N5ZS2TFT.

Fox, D. N., W. J. Teague, C. N. Barron, M. R. Carnes, and C. M. Lee. The Modular Ocean Data Assimilation System (MODAS) (2002), J. Atmos.Oceanic Technol., 19(2) 240–252, doi:10.1175/1520-0426(2002)019< 0240:TMODAS>2.0.CO;2.

Gade, K. A (2010), Non-singular horizonal position representation. J. Navig., 63, 395–417, doi:10.1017/S0373463309990415.Gao, G., C. Chen, J. Qi, and R. Beardsley (2011), An unstructured-grid, finite-volume sea ice model: Development, validation, and applica-

tion, J. Geophys. Res., 116, C00D04, doi:10.1029/2010JC006688.Helfrich, S. R., D. McNamara, B. H. Ramsay, T. Baldwin, and T. Kasheta (2007), Enhancements to, and forthcoming developments in the Inter-

active Multisensor Snow and Ice Mapping System (IMS), Hydrol. Processes, 21(12), 1576–1586, doi:10.1002/hyp.6720.Hibler, W. D. (1979), Dynamic thermodynamic sea ice model, J. Phys. Oceanogr., 9(4), 815–846, doi:10.1175/1520-0485(1979)009< 0815:

ADTSIM>2.0.CO;2.Hibler, W. D. (1980), Modeling a variable thickness sea ice cover, Mon. Weather Rev., 108(12), 1943–1973, doi:10.1175/1520-

0493(1980)108< 1943:MAVTSI>2.0.CO;2.Hill, C., C. DeLuca, Balaji, M. Suarez, and A. da Silva (2004), The architecture of the earth system modeling framework, Comput. Sci. Eng.,

6(1), 18–28, doi:10.1109/MCISE.2004.1255817.Hogan, T. F., T. E. Rosmond, and R. Gelaro (1991), The description of the navy operational global atmospheric prediction system, NOARL

Rep. 13, Naval Res. Lab., Stennis Space Cent., Hancock, Miss.Hogan, T. F., et al. (2014), The Navy Global Environmental Model, Oceanography, 27(3), 116–125.Hollinger, J. P. (1991), DMSP Special Sensor Microwave Imager Calibration/Validation––Final Report Volume II, Naval Res. Lab., Washington, D. C.Hollingsworth, A., K. Arpe, M. Tiedtke, M. Capaldo, and H. Savijarvi (1980), The performance of a medium-range forecast model in winter-

impact of physical parameterizations, Mon. Weather Rev., 108(11), 1736–1773, doi:10.1175/1520-0493(1980)108< 1736:TPOAMR>2.0.CO;2.Hunke, E., and W. Lipscomb (2008), CICE: The Los Alamos Sea Ice Model documentation and software user’s manual version 4.0, Tech. Rep.

LA-CC-06-012, Los Alamos Natl. Lab., Los Alamos, N. M.Hunke, E. C., and J. K. Dukowicz (1997), An elastic-viscous-plastic model for sea ice dynamics, J. Phys. Oceanogr., 27(9), 1849–1867, doi:

10.1175/1520-0485(1997)027< 1849:aevpmf>2.0.co;2.Hunke, E. C., and J. K. Dukowicz (2002), The elastic-viscous-plastic sea ice dynamics model in general orthogonal curvilinear coordinates on

a sphere-incorporation of metric terms, Mon. Weather Rev., 130(7), 1848–1865, doi:10.1175/1520-0493(2002)130< 1848:TEVPSI>2.0.CO;2.Jay, C., B. Marcot, and D. Douglas (2011), Projected status of the Pacific walrus (Odobenus rosmarus divergens) in the twenty-first century,

Polar Biol., 34(7), 1065–1084, doi:10.1007/s00300-011-0967-4.Johannessen, O., et al. (2004), Arctic climate change: Observed and modelled temperature and sea-ice variability, Tellus, Ser. A, 56(4),

328–341, doi:10.1111/j.1600-0870.2004.00060.x.Kay, J. E., et al. (2015), The Community Earth System Model (CESM) Large Ensemble Project: A community resource for studying climate

change in the presence of internal climate variability, Bull. Am. Meteorol. Soc., 8, 1333–1349, doi:10.1175/BAMS-D-13-00255.1.Krishfield, R., J. Toole, A. Proshutinsky, and M. L. Timmermans (2008), Automated ice-tethered profilers for seawater observations under

pack ice in all seasons, J. Atmos. Oceanic Technol., 25(11), 2091–2105, doi:10.1175/2008jtecho587.1.Large, W. G., J. C. Mcwilliams, and S. C. Doney (1994), Oceanic vertical mixing––A review and a model with a nonlocal boundary-layer

parameterization, Rev. Geophys., 32(4), 363–403, doi:10.1029/94RG01872.Ledoux, H., and C. Gold (2005), An efficient natural neighbour interpolation algorithm for geoscientific modelling, in Developments in

Spatial Data Handling, edited by P. F. Fisher, pp. 97–108, Springer, Berlin, doi:10.1007/3-540-26772-7_8.Lemieux, J.-F., et al. (2015), The Regional Ice Prediction System (RIPS): Verification of forecast sea ice concentration, Q. J. R. Meteorol. Soc.,

doi:10.1002/qj.2526, in press.Lipscomb, W. (2001), Remapping the thickness distribution in sea ice models, J. Geophys. Res., 106(C7), 13,989–14,000, doi:10.1029/

2000JC000518.Markus, T., and D. J. Cavalieri (2000), And enhancement of the NASA Team sea ice algorithm, IEEE Trans. Geosci. Remote Sens., 38(3),

1387–1398, doi:10.1109/36.843033.Markus, T., and D. J. Cavalieri (2009), The AMSR-E NT2 sea ice concentration algorithm: Its basis and implementation, J. Remote Sens. Soc.

Jpn., 29(1), 216–225.

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 18

Maslowski, W., D. Marble, W. Walczowski, U. Schauer, J. Clement, and A. Semtner (2004), On climatological mass, heat, and salt transportsthrough the Barents Sea and Fram Strait from a pan-Arctic coupled ice-ocean model simulation, J. Geophys. Res., 109, C03032, doi:10.1029/2001JC001039.

McGeehan, T., and W. Maslowski (2012), Evaluation and control mechanisms of volume and freshwater export through the Canadian ArcticArchipelago in a high-resolution pan-Arctic ice-ocean model, J. Geophys. Res., 117, C00D14, doi:10.1029/2011JC007261.

Meier, W., F. Fetterer, J. S. Stewart, and S. Helfrich (2015), How do sea ice concentrations from operational data compare with passivemicrowave estimates?, Ann. Glaciol., 56(69), 332–340, doi:10.3189/2015AoG69A694.

Metzger, E., et al. (2014), US Navy operational global ocean and Arctic ice prediction systems, Oceanography, 27(3), 32–43.Metzger, E. J., H. E. Hurlburt, A. J. Wallcraft, J. F. Shriver, O. M. Smedstad, P. Thoppil, and D. S. Franklin (2008), Validation test report for the

Global Ocean Prediction System v3.0 - 1/128 HYCOM/NCODA phase 1. Tech. Rep. NRL/MR/7320-0809148, Naval Res. Lab., Stennis SpaceCent., Hancock, Miss.

Metzger, E. J., A. J. Wallcraft, P. G. Posey, O. M. Smedstad, and D. S. Franklin (2013), The switchover from NOGAPS to NAVGEM 1.1 atmos-pheric forcing in GOFS and ACNFS. Tech. Rep. NRL/MR/7320-13-9486, Naval Res. Lab., Stennis Space Cent., Hancock, Miss.

Metzger, E. J., P. G. Posey, P. G. Thoppil, T. L. Townsend, A. J. Wallcraft. (2015), Validation test report for the Global Ocean Forecast System3.1-1/128 HYCOM/NCODA/CICE/ISOP Tech. Rep NRL/NR/7320-15-9579, Naval Res. Lab., Stennis Space Cent., Hancock, Miss.

Murphy, A. H., and E. S. Epstein (1989), Skill scores and correlation-coefficients in model verification, Mon. Weather Rev., 117(3), 572–581,doi:10.1175/1520-0493(1989)117< 0572:SSACCI>2.0.CO;2.

NIC (2008), IMS Daily Northern Hemisphere Snow and Ice Analysis at 1 km, 4 km, and 24 km Resolutions, Natl. Snow and Ice Data Cent.,Boulder, Colo., doi:10.7265/N52R3PMC.

NIMA (2000), Department of Defense World Geodetic System 1984: Its definition and relationships with local geodetic systems. Tech. Rep.TR8350.2, National Imagery and Mapping Agency, St. Louis, Mo.

Posey, P. G., E. J. Metzger, A. J. Wallcraft, R. H. Preller, O. M. Smedstad, and M. W. Phelps (2010), Validation of the 1/128 Arctic Cap Nowcast/Forecast System (ACNFS), Tech. Rep. NRL/MR/7320-10-9287, Naval Res. Lab., Stennis Space Centre, Mass.

Posey, P. G., E. J. Metzger, A. J. Wallcraft, D. A. Hebert, R. A. Allard, O. M. Smedstad, M. W. Phelps, F. Fetterer, J. S. Stewart, S. R. Helfrich(2015), Improving Arctic sea ice edge forecasts by assimilating high horizontal resolution sea ice concentration data into the U.S. Navy’sice forecast systems, Cryosphere, 9, 1–11, doi:10.5194/tc-9-1-2015.

Preller, R. H., P. G. Posey, W. Maslowski, D. Stark and T. Pham (2002), Navy sea ice prediction systems, Oceanography, 15(1), 44–56.Rasmussen, T., N. Kliem, and E. Kaas (2010), Modelling the sea ice in the Nares Strait, Ocean Modell., 35(3), 161–172, doi:10.1016/

j.ocemod.2010.07.003.Sakov, P., F. Counillon, L. Bertino, K. A. Lisæter, P. R. Oke, and A. Korablev (2012), TOPAZ4: An ocean-sea ice data assimilation system for the

North Atlantic and Arctic, Ocean Sci., 8(4), 633–656, doi:10.5194/os-8-633-2012.Scholz, P., G. Lohmann, Q. Wang, and S. Danilov (2013), Evaluation of a Finite-Element Sea-Ice Ocean Model (FESOM) set-up to study the

interannual to decadal variability in the deep-water formation rates, Ocean Dyn., 63(4), 347–370, doi:10.1007/s10236-012-0590-0.Schweiger, A., and J. Zhang (2015), Accuracy of short-term sea ice drift forecasts using a coupled ice-ocean model, J. Geophys. Res. Oceans.,

doi:10.1002/2015JC011273 (accepted, in press).Schweiger, A., R. Lindsay, J. Zhang, M. Steele, H. Stern, and R. Kwok (2011), Uncertainty in modeled Arctic sea ice volume, J. Geophys. Res.,

116, C00D06, doi:10.1029/2011JC007084.Serreze, M., M. Holland, and J. Stroeve (2007), Perspectives on the Arctic’s shrinking sea-ice cover, Science, 315(5818), 1533–1536, doi:

10.1126/science.1139426.Serreze, M. C., and J. Stroeve (2015), Arctic sea ice trends, variability, and implications for seasonal ice forecasting. Philos. Trans. R. Soc. A,

373, 1–16, doi:10.1098/rsta.2014.0159.Smith, G. C., et al. (2015), Sea ice forecast and verification in the Canadian Global Ice Ocean Prediction System, Q. J. R. Meteorol. Soc., doi:

10.1002/qj.2555. (in press).Snyder, J. P. (1987), Map projections––A working manual, Tech. Rep. 1395, U.S. Gov. Print. Off., Washington, D. C., ISBN: 87-ti00250. [Avail-

able at http://pubs.usgs.gov/pp/1395/report.pdf, last accessed 2 Oct. 2015.]Stammer, D., C. Wunsch, R. Giering, C. Eckert, P. Heimbach, J. Marotzke, A. Adcroft, C. Hill, and J. Marshall (2002), Global ocean circulation during

1992-1997, estimated from ocean observations and a general circulation model, J. Geophys. Res., 107(C9), 3118, doi:10.1029/2001JC000888.Toole, J. M., R. A. Krishfield, M. L. Timmermans, and A. Proshutinsky (2011), The ice-tethered profiler: ARGO of the Arctic, Oceanography,

24(3), 126–135.U.S. Navy (2014), U.S. Navy Arctic Roadmap 2014-2030, CreateSpace Publishing, Washington, D. C.van Scheltinga, A., P. Myers, and J. Pietrzak (2010), A finite element sea ice model of the Canadian Arctic Archipelago, Ocean Dyn., 60(6),

1539–1558, doi:10.1007/s10236-010-0356-5.Van Woert, M., C. Zou, W. Meier, P. Hovey, R. Preller, and P. Posey (2004), Forecast verification of the Polar Ice Prediction System (PIPS) sea

ice concentration fields, J. Atmos. Oceanic Technol., 21(6), 944–957, doi:10.1175/1520-0426(2004)021< 0944:FVOTPI>2.0.CO;2.Wunsch, C., and P. Heimbach (2013), Dynamically and kinematicallly consistent global ocean circulation and ice state estimates. in Ocean

Circulation and Climate: A 21 Century Perspective. edited by G. Siedler, et al., Intern. Geophys., vol. 103, pp. 553–579, Academic Press,Oxford, U. K.

Zhang, J., and R. Lindsay, A. Schweiger, and I. Rigor (2012), Recent changes in the dynamic properties of declining Arctic Sea Ice: A modelstudy, Geopohys. Res. Lett., 39, L20503, doi:10.1029/2012GL053545.

Zhang, J., and D. Rothrock (2003), Modeling global sea ice with a thickness and enthalpy distribution model in generalized curvilinear coor-dinates, Mon. Weather Rev., 131(5), 845–861, doi:10.1175/1520-0493(2003)131< 0845:MGSIWA>2.0.CO;2.

Journal of Geophysical Research: Oceans 10.1002/2015JC011283

HEBERT ET AL. U.S. NAVY SHORT-TERM SEA ICE FORECASTING 19